List your company
ResearchFormerly WizWand

Scene Text Understanding on OCRBench

804ScoreInternVL2.5
-32.1392184.9354402.01619.0846Dec 22, 2025Jan 23, 2026Feb 24, 2026Mar 28, 2026Apr 29, 2026May 31, 2026Jul 2, 2026
Updated 3mo ago

Evaluation results

MethodLinks
InternVL2.5# train tokens=0.5T, Architecture Category=Token Insertion – Proprietary2025.12804
Qwen2.5-VL 3B (reproduced)Res.=≤ 896²2025.12804
Qwen2.5-VL 3B (reported)Res.=Native2025.12797
CASA⊕ Qwen2.5-VLRes.=≤ 896²2025.12790
VideoLLaMA3Architecture Category=Token Insertion – Proprietary2025.12779
Qwen2-VL# train tokens=1.4T, Architecture Category=Token Insertion – Proprietary2025.12767
SmolVLMArchitecture Category=Token Insertion – Public data, LLM Size=2B2025.12729
InsertionHe-2B# train tokens=0.1T, Architecture Category=Token Insertion – Public data, LLM Size=2B2025.12728
CASA⊕# train tokens=0.3T, Architecture=CASAHe-2B, CASA design variant=⊕, LLM Size=2B2025.12728
CASA→# train tokens=0.3T, Architecture=CASAHe-2B, CASA design variant=→, LLM Size=2B2025.12723
CASA∨# train tokens=0.3T, Architecture=CASAHe-2B, CASA design variant=∨, LLM Size=2B2025.12694
mPLUG-Owl3 8B# train tokens=0.1T, Architecture Category=Cross-attention-based – Public data, LLM Size=8B2025.12527
mPLUG-Owl3 2B# train tokens=0.1T, Architecture Category=Cross-attention-based – Public data, LLM Size=2B2025.12450
InternVL3.5-38BAttack=Benign2026.070.866
Qwen2.5-VL-72B-InstructAttack=Benign2026.070.84
Qwen2.5-VL-7B-InstructAttack=Benign2026.070.837
InternVL3.5-8BAttack=Benign2026.070.827
InternVL3.5-4BAttack=Benign2026.070.813
Optimization-based Token SelectionTarget Model=InternVL3.5-38B2026.070.812
Qwen2.5-VL-3B-InstructAttack=Benign2026.070.789
Permutation AttackTarget Model=Qwen2.5-VL-3B-Instruct, Token Manipulation Ratio=0.12026.070.776
Masking AttackTarget Model=Qwen2.5-VL-3B-Instruct, Token Manipulation Ratio=0.12026.070.735
Gaussian Perturbation AttackTarget Model=Qwen2.5-VL-3B-Instruct, Token Manipulation Ratio=0.12026.070.668
Optimization-based Token SelectionTarget Model=InternVL3.5-8B2026.070.655
VTM-Attack (Naïve Average)Target Model=Qwen2.5-VL-3B-Instruct2026.070.634
Sign Flip AttackTarget Model=Qwen2.5-VL-3B-Instruct, Token Manipulation Ratio=0.12026.070.357
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-3B-Instruct, Base Attack=Sign Flip, Iterations=30, Learning Rate=0.052026.070.305
Optimization-based Token SelectionTarget Model=InternVL3.5-4B2026.070.157
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-7B-Instruct2026.070.15
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-72B-Instruct2026.070.02